Speaker identification: The case for speech vector analysis.
Harry Hollien, James D. Harnsberger · The Journal of the Acoustical Society of America · 2010
The problem of identifying speakers from voice analysis is a serious one. Many analytical procedures have been proposed; most have been based on signal analysis algorithms. Yet it is clear that human perceivers can make accurate identifications (even under difficult circumstances), and extensive research supports this position. How do they accomplish this task? They do so by acoustic and temporal assessment of the talker’s speech/voice signal. Several analysis procedures have attempted to mimic human perception for this purpose, and a brief review of the relevant data will be presented. It will be followed by a presentation of the results of a four-vector experiment in which each vector integrates 3–5 speech parameters. Identification of 28 male voices resulted from three replications in a field of 10 foil voices. It was found that identification scores for the voice, vowel, and fundamental frequency vectors were high, and that for the temporal vector was modest. Moreover, of the summation scores across all vectors and replications, only one failed to identify the target speaker. While these results were based on high-quality audio recordings, it can still be argued that robust speaker identification is possible from a limited set of speech and voice characteristics.